GenSearcher/Gen-Searcher-8B

Gen-Searcher-8B: A Deep Research Agent for Image Generation

Gen-Searcher-8B is an 8 billion parameter multimodal deep research agent, representing the first attempt to train such an agent specifically for image generation tasks that demand complex real-world knowledge. Developed by Kaituo Feng and his team, this model is designed to enhance the accuracy and relevance of image synthesis by performing extensive pre-generation research.

Key Capabilities

• Agentic Search: Capable of searching the web to gather information and browse evidence.
• Multi-Source Reasoning: Reasons over multiple sources of information to inform image generation.
• Visual Reference Search: Searches for visual references to guide the synthesis process.
• Enhanced Accuracy: Enables more accurate and up-to-date image generation in real-world scenarios.
• Benchmark Performance: Achieves significant improvements, with over 15-point gains on the KnowGen and WISE benchmarks.
• Transferability: Demonstrates strong transferability across various image generators.

Training and Resources

The model was trained using two dedicated datasets: Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k. A new benchmark, KnowGen, was also introduced for search-grounded image generation. All code, models, data, and benchmarks are fully released and available on the GitHub Repository and detailed in the accompanying paper.

Good For

• Applications requiring image generation grounded in real-world, up-to-date information.
• Developers looking for a model that can perform deep research before synthesizing images.
• Research into agentic AI for multimodal tasks and complex knowledge integration.